My original question asked how to balance outgoing traffic between two
ethernet interfaces (on one machine) plugged into a switch. Both
interfaces were on the same subnet. I also had referred to a document
that is now MIA (http://www.grandjunction.com/custsupp/techtips/sunserv.html).
Cisco bought grandjunction, and it might be on their website now.
Many thanks to all who replied, and especially Gene Rackow and Karlheinz
Pischke, who, despite my poorly worded e-mail, knew just what I was trying
to do (grin).
In short, I'd setup the two interfaces with unique MAC addresses,
and two unique IP addresses. Then I setup round-robin DNS so that every
time a machine did a lookup on our server (redhook) the IP address that
was returned alternated between the two interfaces (.1 and .16 in my
case). This split up redhook's incoming traffic between the two
interfaces nicely.
But outgoing traffic was still all going out on the primary interface.
Gene's reply hit the nail on the head. I either need to split my net
into two subnets, assign an IP from the second subnet to the second
interface (both of my interfaces had IP's less than 128) and setup
the broadcast and netmask addresses appropriately. Alternatively, I
could build static routes to tell the server to use the second interface
for traffic to certain machines.
Neither method automatically balances the load like round-robin DNS
does for the incoming traffic, but it's still better than only using one
interface for all outgoing traffic (grin). Now, just using a few static
routes (a quick hack 'till I can subnet and re-assign IPs) I'm getting
closer to 20 mbit/s throughput both into and out of our server.
Brent Bice
bbice@persistence.com
> From: Gene Rackow <rackow@mcs.anl.gov>
>
> [ Mentioned Cisco buying GrandJunction ]
> Now that you have the machines set so that inbound traffic can get to either
> interface, you need to get outbound set the same way. To do that you need
> to convince the sparc that the best route to a host is not via the first
> interface. You could do that by munging routing so the sparc thinks there
> are 2 seperate subnets where the rest of the system only claims one, or
> you can specify individual host routes to go to the send port.
>
> Method 1 is doing for each machine you want to talk to via port 2
> route add machine second.ip#
> This allows you to tune things to balance the load on a machine by machine
> basis, but can get quite consuming.
>
> Method 2 is basicly changing the ifconfig line for the 2 ports at
> boot time to specify a tighter netmask. If you have a number of subnets,
> things will get really ugly at this point. On a single net, it's not bad.
> This keeps things simple in that you balance load depending on upper/lower
> half of your network.
>
> Note that neither method will give you a load balancing if several machine
> you have chosen as part of set 1 all want to talk at the same time to the
> server and none of set 2 are active.
>
> -_Gene
>
> From: Matthew Stier - BSG Corporation <matthew.stier@imonics.com>
>
> Knowing what OS release you are working with, would help in defining what
> needs to be done.
>
> It sounds like you have two network interfaces attached to the same physical
> network. Can I assume that you have taken the necessary steps to change
> the second interfaces MAC address?
>
> If what your trying to do, is balance outbound traffic between the two
> interfaces, good luck. The IP sections of the kernel simply care to
> find a way out, they were written to balance load across several physical
> interfaces. What is probably happening, is that the kernel is trying to
> route the outbound packet, and lacking a specific route, is trying to find
> the first interface that will get it there, and for you, that would be le0.
>
> The only real answer, is replacing your 10Base-? interface with a faster
> technology, such as FDDI or 100Base-T.
>
> Note: Cisco purchase GrandJunction. Have you tried Cisco's website?
> From: Rasana Atreya <Rasana.Atreya@library.ucsf.edu>
>
> I'm not sure you'll find it here, but check out:
> ftp://ftp.picarefy.com/pub/Sun-Hardware-Ref/
>
> Rasana
>
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> ~ Rasana Atreya Voice: (415) 476-3623 ~
> ~ System Administrator Fax: (415) 476-4653 ~
> ~ Library & Ctr for Knowledge Mgnt, Univ. of California at San Francisco ~
> ~ 530 Parnassus Ave, Box 0840, San Francisco, CA 94143-0840 ~
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>
> From: kpischke@cadence.com (Karlheinz Pischke)
>
> The problem is the ROUTING
>
> You have not setup the routing tables for which interface should make conncetion
> to which machine. So always the first entry in the routing table is used and
> this is le0 ! Clear all outgoing traffic goes to interface le0.
>
> Solutions:
>
> 1) static routing table with an entry for each host
> Static routing tables are really ugly to manage !!
> So forget about it if you have many (> 10) machines or lot of changes.
>
> 2) split your network in smaller parts SUBNETTING
> You have only once a big job to configure all machines to a (probably)
> new internet address and netmask.
>
> You may get a problem with your router(s). All subnets "around" the router(s)
> should have the SAME netmask. [you will have to ask a router specialist for
> detail]
>
> 3) the only NFS server (I know of) which is able to respond automatically
> to the same interface where the request came in.
> So I suggest using an Auspex server. (Price category UltraServer 8000)
>
> btw: I'm evaluating an Auspex server for a while. I'm a fan of this machine.
>
> Mit freundlichen Gruessen / Best Regards
>
> Karlheinz Pischke
>
> Karlheinz Pischke, Dipl.-Inf., Applikationsingenieur
>
> Email: kpischke@cadence.com C A D E N C E GmbH
> Richard-Reitzner-Allee 8 (H 2)
> Tel.: 49 - 89 - 4563-0 85551 Haar
> Fax: 49 - 89 - 4563-1919 Germany
> From: sozoa@atmel.com (Steve Ozoa)
>
> Grand Junction was bought by Cisco some time ago. You may have some luck under
> Cisco's home page at http://www.cisco.com.
>
> If not, I think that's the same tip sheet I downloaded a few months ago. If
> you don't get it another way, I can fax it to you.
>
> From: Bert Shure <Bert_Shure%SOLSOURCE@notes.worldcom.com>
>
> cisco bought grand junction. you might try www.cisco.com.
>
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:11:09 CDT